use a tuple instead of Ref in broadcasts

[rafaqz]

Ref is not free

[jw3126]

LGTM. In my experience the the compiler optimizes broadcast Ref like this very well. Are you aware of an example where Tuple is acually faster in a broadcast like this? Anyway I like simple stupid Tuple more than relying on optimizer so LGTM.

[rafaqz]

Benchmarking something basic was a Ns faster even though the allocation was elided in Ref. I don't have the code right now to demonstrate.

And yes keeping it simple is better anyway, this is really low level code.

[aplavin]

Wow good to know if that's the case!
I've always thought that Ref is the recommended solution for broadcasting as a scalar – it's even recommended in the Julia docs. Do you happen to have a relevant Julia issue that Ref is slower? Everyone else also deserves to know :)

[rafaqz]

I thought performance was not the reason for the Ref recommendation, instead it's because of scalar unwrapping details.

Tuple logically cant be slower than Ref, and Ref needs compiler optimisations to match a Tuple. I don't think this is any huge revelation.

See e.g. https://discourse.julialang.org/t/marking-types-as-scalar-for-broadcasting-ref-vs-tuple/29105

[rafaqz]

Also see the original issues for performance concerns.

JuliaLang/julia#18379

The compiler has to actively elide Ref allocations. My suspicion is at times this complexity will stop it seeing other optimisations in a larger context.

That's important here where this small broadcast will often be in the middle of a much more complicated process that needs high performance.

[rafaqz]

Here's a very specific reference to the problem:

JuliaLang/julia#39151 (comment)

And here, Ref blocking constant propagation is identified as a key problem. That may be the reason my own benchmarks were slower too:

JuliaLang/julia#39151 (comment)